A Unsupervised graph comparison learning-based click-through rate prediction model

Abstract Click-through rate (CTR) prediction refers to the estimation of the probability that a given user will click on a given advertisement, with the objective of displaying items of greater interest to the user. In this paper, we propose a clickthrough rate prediction model based on unsupervised graph comparison learning (USGCL-CTR) to address the issue of a decline in model prediction accuracy due to the neglecting the introduction of noise caused by useless feature interactions in existing CTR prediction models based on GNNs. To facilitate the transformation of the modeling task of feature interactions into the modeling of node interactions on the corresponding graph, the feature domain is visually represented as a graph structure following the embedding layer. Subsequently, noisy edges are removed by a designed feature interaction selection layer based on unsupervised graph comparison learning. In USGCL-CTR, we utilize metric learning to establish an ’anchor graph’, serving as a guide for the optimization of the graph structure. We then apply contrast loss to maximize the interaction information between the anchor graph and the learned structure. By eliminating unproductive feature interactions, the denoised feature map significantly enhances the precision of CTR predictions. Comparative experiments were carried out against eight baseline models on three open-source datasets. The results clearly showed that our proposed USGCL-CTR model surpassed all baseline models in performance.